| name | session-handoff |
| description | End-of-session handoff that captures session knowledge, dispatches output across the canonical 7-bucket docs/ taxonomy (decisions/runbooks/analysis/references/reviews/handoffs/deliverables — aligned with memory-hygiene v3.3), triggers a doc-freshness reverse-lint + skill-freshness audit to catch stale normative guidance, emits the future-to-do plan's follow-up items as GitHub issues, updates memory, and prepares next-session prompts. Use when: (1) user says 'wrap up', 'hand over', 'create handoff', 'end of session', 'write handoff', 'session handoff'; (2) non-trivial work session (3+ tasks) is ending; (3) context window is approaching limits; (4) user says 'consolidate', 'what's the current state', 'start here document' after parallel sessions; (5) the session produced artifacts that belong in more than one docs/ bucket (ADR + analysis + runbook + review). Includes cross-session consolidation when 3+ handoffs accumulate and a mandatory reverse-lint verify step against any lessons.md / feedback_*.md touched this session. |
| version | 1.9.1 |
| triggers | ["wrap up","session handoff","end of session","create handoff","hand over","write handoff","consolidate","consolidate handoffs","what's the current state","start here document"] |
Session Handoff v1.9 — Bucket-aware + reverse-lint + skill-freshness + issue emission
Comprehensive end-of-session knowledge capture with built-in cross-session
consolidation. Ensures nothing is lost between sessions and produces a single
source of truth when multiple handoffs accumulate.
v1.9 alignment with memory-hygiene v3.3: session output is dispatched across
the canonical 7-bucket docs/ taxonomy — not just docs/handoffs/. At the end of
the workflow, invokes doc-freshness-reverse-lint and a skill-freshness audit
against any memory or SKILL.md files touched this session to surface stale normative
guidance, and emits the future-to-do plan's follow-up items as GitHub issues so
nothing relies on a future session re-reading the handoff to act on them.
Counterpart skill: memory-hygiene v3.3 cleans Claude's persistent memory +
audits project docs/ against the same taxonomy. Run memory-hygiene after
10+ sessions or when docs/ has drifted.
When to use
- End of any non-trivial work session (3+ tasks completed)
- User says "wrap up", "hand over", "create handoff", or similar
- Before context window approaches limits
- After parallel sessions complete and you need one "start here" document
- User says "consolidate", "what's the current state"
Canonical 7-bucket docs/ taxonomy (from memory-hygiene v3.3)
Session output is dispatched to the right bucket — not dumped into one handoff file.
A typical rich session produces artifacts in 3-5 of these 7 buckets simultaneously.
| # | Bucket | Write here when the session produced... | Filename convention |
|---|
| 1 | docs/decisions/ | An architectural / methodological choice worth preserving (go/no-go, tradeoff, supersession) | NNNN-kebab-case.md (ADR — check for duplicate numbers first) |
| 2 | docs/runbooks/ | A new rerun / retrain / operational procedure, or updated steps to an existing one | <verb>_<noun>.md (e.g. retrain_propensity.md, rerun_guide.md) |
| 3 | docs/analysis/ | Findings, investigations, diagnostics, discovery write-ups, exploratory analyses | analysis_<topic>.md, discovery_<topic>.md, findings_<topic>.md |
| 4 | docs/references/ | New/updated schema, data dictionary, API ref, project-convention doc | <system>_reference.md, data_dictionary.md, <topic>_dictionary.md |
| 5 | docs/reviews/ | Review-panel output, peer review, audit report | review_<topic>.md, <topic>_audit_report.md, next_stage_<topic>.md |
| 6 | docs/handoffs/ | Always — this session's handoff doc + next-session prompt + (optional) parallel prompts | session_N_handoff.md, session_N+1_prompt.md, session_N+1b_<topic>_prompt.md |
| 7 | docs/deliverables/ | External-facing artifact (client draft, published output, slide deck, PDF, XLSX) | Keep original extension; add a .provenance.md sibling if the artifact was generated |
Reserved top-level file: only docs/README.md. No other loose files at docs/*.
Exclude from docs/ entirely: __pycache__/, .py scripts (belong in scripts/), .DS_Store.
If a session artifact doesn't fit any bucket, surface it to the user — don't invent an 8th bucket.
Edge cases
- No git repo: Skip git log and branch status steps. Note this in the handoff doc.
- No
memory/ directory: Create it. Initialize lessons.md and sessions_archive.md.
- No
docs/ directory: Create docs/handoffs/ and docs/plans/.
- First session (no prior handoffs): Skip consolidation. Write session_1_handoff.md.
- Single handoff exists: Skip consolidation — it only adds value with 3+ handoffs.
- Handoff docs use different naming: Scan for
session*handoff* and *_handoff.md patterns.
The Checklist (execute in order)
Phase 0: Label audit (loose mode, blocking-with-escape-hatch)
Before emitting any handoff doc that contains code → human-label tables (e.g.
AD — Accepted Fully Salesforce status codes, HTTP-status legends, enum
descriptions), run the label audit. This is the author-side gate that
prevents the "predecessor handoff fabricates semantic labels" failure mode
(see ~/.claude/lessons.md L-S109b-1 and axioms.md § Authoritative Labels).
Run after Phase 1 Step 2 (handoff doc draft) and before any other phase.
LABEL_AUDIT="${CLAUDE_PLUGIN_ROOT:+${CLAUDE_PLUGIN_ROOT}/scripts/label_audit.py}"
[ -f "$LABEL_AUDIT" ] || LABEL_AUDIT="$HOME/.claude/skills/session-handoff/scripts/label_audit.py"
python3 "$LABEL_AUDIT" docs/handoffs/session_N_handoff.md
If the script is not found at either location, log "label audit: not installed —
labels unverified" in the handoff doc and continue — never block the handoff on
missing tooling.
Behavior:
- Exit 0 (clean) — no code-legend rows detected, OR every flagged row
carries an inline tag. Continue.
- Exit 1 (blocking — untagged rows) — print the offending rows, then
fix one of these ways before continuing:
- Add
[verified: <repo-relative-path>:<line>] after the description for
every row whose label has been confirmed against an authoritative source
(vendor doc, INFORMATION_SCHEMA, internal data dictionary derived from
real probes). Cite the file + line that contains the verification.
- Add
[HYPOTHESIS] after the description for any row whose label is a
guess, so the receiving session knows to re-probe.
- If the table is a retrospective reference (e.g., quoting an old
fabricated table to explain why the new system exists) and re-tagging is
onerous, set frontmatter
label-audit-skipped: <reason> to bypass.
- Exit 2 (skipped via frontmatter) — print the skip reason; the
receiving session sees it and knows to treat ALL labels in this doc as
unverified.
The escape hatch is a load-bearing feature, not a workaround. False
positives (e.g., a git-commit-hash table with 3-letter codes) cost 30s of
human ack to bypass; false negatives cost a fabricated label shipping to a
client. The asymmetry is by design.
Run the same scanner against any next-session prompts you write in
Phase 3 (session_N+1_prompt.md, parallel prompts) — the receiving
session is downstream of these too.
Phase 1: Capture (what happened)
-
Scan git log for all commits this session:
git log --oneline --since="today" | head -30
-
Write handoff doc -> docs/handoffs/session_N_handoff.md
- Section 1: What was completed (with specifics — PRs, test counts, key metrics)
- Section 2: What remains (prioritised, with file pointers)
- Section 3: Blockers & open issues (what's stuck, what's waiting on external input, what failed)
- Section 4: Key decisions (table: Decision | Resolution | Rationale)
- Section 5: Files modified (table)
- Section 6: Branch status
-
Scan for missed lessons — review the session for:
- Debugging that required non-obvious investigation
- Bugs with root causes worth documenting
- Patterns that would help in future similar situations
- User corrections that should become rules
Add to
memory/lessons.md with sequential numbering.
-
Create/update memory feedback files for significant findings:
- New feedback files for major discoveries (
feedback_*.md)
- Update
reference_*.md if project constants changed
- Update architectural decision records with new decisions
Phase 2: Dispatch to the 7 canonical buckets
Before writing the handoff doc (Phase 1 step 2), triage everything the session produced and
decide which of the 7 buckets each artifact belongs to. The handoff doc then cross-links each
bucket output rather than duplicating its content.
-
Triage session output. For each significant artifact the session produced, pick a bucket
using the taxonomy table above. If unsure between two buckets, prefer the more specific one
(e.g., a review of an analysis → reviews/ not analysis/).
-
docs/decisions/ — write/update ADRs for significant architectural/methodological decisions.
- Path:
docs/decisions/NNNN-kebab-case.md
- Check for duplicate ADR numbers first (grep
^\d{4}- in docs/decisions/)
- Status, Context, Decision, Consequences, Confirmation, (optional) Supersedes/SupersededBy
-
docs/runbooks/ — capture new or updated operational procedures.
- New rerun/retrain procedure → new file under
docs/runbooks/
- Updated steps to existing runbook → edit in place, add a dated changelog entry at the top
-
docs/analysis/ — write up findings and investigations.
- Exploratory analysis, diagnostic runs, discovery write-ups
- Include methodology + what-would-change-the-conclusion so future sessions can re-evaluate
-
docs/references/ — update schemas, data dictionaries, convention docs.
- If the session added columns to a table, update the data dictionary
- If the session changed project conventions, update the relevant reference doc
-
docs/reviews/ — write review-panel or audit output.
- Review reports, peer reviews, audit findings produced this session
-
docs/handoffs/ — always write session handoff + next prompt (Phase 3 below covers this).
-
docs/deliverables/ — record external-facing artifacts.
- Client drafts, published outputs, slide decks, PDFs, XLSXs
- If the artifact was generated (not hand-authored), add a
.provenance.md sibling with
source inputs, generation date, and regeneration command
-
Update future plan -> docs/plans/future_sessions_plan.md
- Mark completed items as DONE
- Add new items discovered during session
- Update status of in-progress items
-
Update roadmap (if it exists)
-
Update sessions archive -> memory/sessions_archive.md
- Add entry with session number, date, one-line summary, key outcomes
- Include a bucket-footprint line (e.g. "Wrote to: decisions/, analysis/, handoffs/")
-
Update MEMORY.md index
- Add new memory files
- Update lesson count
- Add session reference
Emitting follow-up items as GitHub issues
This is the issue-filing half of step 13. The future-to-do plan's whole job is
preserving context across sessions — but a follow-up that lives only as plan
prose decays into a stale TODO, because the next session has no actionable
breadcrumb. Filing each follow-up as a GitHub issue closes that loop.
Which items to emit: the new follow-up items this session discovered and is
not completing now — whatever the plan calls them (Next steps, Follow-ups,
Deferred, Open questions). Skip items already marked DONE and items already
tracked by an existing issue.
Default behavior — dry-run preview. Do NOT file issues silently. For each
follow-up item, draft a gh issue create payload and show the user the full
command(s) for inspection first:
gh issue create \
--repo <owner>/<name> \
--title "<concise follow-up title>" \
--body "<why this matters + origin session/PR + file pointers>" \
--label "follow-up"
Present all drafted commands as one batch, then let the user approve, edit, or
skip individual items. Run the approved gh issue create commands only after the
user confirms. (If the user has explicitly asked for autonomous operation, you
may file directly — but the dry-run preview is the default.)
Target repo resolution:
- If the user named a repo, use it (
--repo owner/name).
- Otherwise auto-detect from the current git remote:
gh repo view --json nameWithOwner -q .nameWithOwner (falls back to parsing
git remote get-url origin).
- If neither resolves, skip issue emission — leave the items as plan text and
note "issue emission skipped: no target repo" in the handoff doc.
De-duplicate before filing. For each drafted title, search the target repo's
open issues so re-running the handoff on the same plan doesn't double-file:
gh issue list --repo <owner>/<name> --state open --search "<title keywords>" \
--json number,title
If a clear title match exists, drop that item from the batch and reference the
existing issue number in the plan instead of filing a duplicate.
After filing: annotate the corresponding plan item with its issue number
(e.g. - [ ] Track 2 engagement-width follow-up — #201) so the plan and the
issue tracker stay linked.
Graceful degradation: if gh is not installed or not authenticated, skip
issue emission, keep the follow-up items as plan text, and note "issue emission
skipped: gh unavailable" in the handoff doc — never block the handoff on it.
Phase 3: Prepare (next session)
-
Write next session prompt -> docs/handoffs/session_N+1_prompt.md
- Key context (what's done, what's blocked)
- Start files to read (include bucket-specific outputs from Phase 2)
- Priority tasks with specific instructions
- Research context
-
Write parallel session prompts when upcoming work can be split into independent streams
When to split: If the next session's scope contains 2+ tasks with zero file overlap, split
them into separate prompts that can run simultaneously on separate branches.
File overlap check: For each task, list the files it will modify. If the sets are disjoint
(e.g., _feature_common/ + dataform/ vs cr_client_dashboard/ only), they're safe to parallelize.
Naming convention: session_N+1_prompt.md (primary) + session_N+1b_[topic]_prompt.md (parallel).
Use a short descriptive suffix like _insight_card_redesign or _cleanup.
Each parallel prompt must include:
- Its own branch name (e.g.,
feat/s71-distance-v6 vs feat/s71b-insight-redesign)
- A "Parallel session" section naming the other stream + confirming no file overlap
- Any shared prerequisites (e.g., "merge PR #N first, then branch from main")
- Its own guardrails (test count, deploy restrictions)
Skip splitting when:
- Tasks share files or have ordering dependencies
- One task is trivially small (< 15 min) — just sequence it
- The user hasn't expressed interest in parallel execution
Phase 4: Commit, PR, and verify (nothing lost)
-
Check for uncommitted changes:
git status --short
git diff --name-only
git log --oneline origin/BRANCH..HEAD
-
Commit all session work — stage and commit in logical groups:
- Code changes first: feature code, bug fixes, tests (one commit with descriptive message)
- Docs second: handoff doc, next-session prompt, plan updates, lessons (separate commit)
- If the session already has multiple commits on a feature branch, add docs as a new commit on the same branch.
- If uncommitted work is on
main, create a feature branch first: git checkout -b feat/sN-description
-
Push and create PR:
git push -u origin <branch-name>
gh pr create --title "feat(sN): <summary>" --body "<PR body with summary + test plan>"
- PR body should include: summary bullets, test plan checklist, line/file counts
- If the session had no code changes (docs-only), use
docs(sN): prefix instead of feat(sN):
-
Review + auto-merge for docs-only handoff PRs. When the PR is purely docs from this skill (handoff doc + next-prompt + maybe a sessions_archive/MEMORY.md edit, no code/tests/deploy), default to: review-with-agent → fix findings → squash-merge. Don't wait for the user to ask. The handoff PR has narrow scope and stalls the loop if it sits open between sessions.
a. Launch a review agent scoped lighter than code review:
- Factual accuracy — cited PR numbers, commit SHAs, issue numbers, tracker IDs all resolve (
gh pr view, gh issue view, git log <sha> -1)
- Dead-reference check — every file path / line number / handoff filename mentioned must resolve in current main, OR be explicitly annotated as living on a non-main branch with a
git show <branch>:<path> recipe. Common trap: predecessor handoff files written on a wip branch that never merged
- Internal consistency — handoff and next-prompt agree on session number, predecessor refs, branch names, what's done vs deferred
- Self-contained next-prompt — receiver can cold-start; no "see above" or unresolved references
- Naming convention — if
b/c suffix used for parallel-session collision, confirm precedent matches and self-references use the suffixed filename
Use voltagent-qa-sec:code-reviewer (or equivalent code-review agent). Request a <250-word report. Surface only HIGH/MEDIUM findings.
b. Address findings via Edit tool. Common patterns:
- Predecessor refs to files on a wip branch → annotate with
git show <branch>:<path> recipe (don't delete the reference; the receiver may want to read it)
- Self-reference bugs (
session_NNN_handoff.md (this) where the actual filename includes a b/c suffix)
module.py:NNN line numbers drift fast — verify against current main or drop the line number
c. Refresh the PR body if findings meaningfully changed the story (note the review-agent pass + what was fixed).
d. Squash-merge: gh pr merge <N> --squash --delete-branch. If --delete-branch fails with "main is already used by worktree at..." (common when running from a worktree that has main checked out elsewhere), the merge still succeeded — clean up the remote ref via gh api -X DELETE repos/<owner>/<repo>/git/refs/heads/<branch>.
e. Sync local main if possible (git checkout main && git pull). Skip silently if the current worktree can't switch (branch checked out elsewhere). Note merge in the final summary table.
For non-docs-only PRs (any code/tests/deploy mixed in): default to leaving the PR open for human review. Only merge if user explicitly asks. Project memory may add a "always run code-reviewer pass before merging PRs" rule — apply that for code PRs.
-
Quick memory hygiene check:
- Any new memory files missing from MEMORY.md?
- Lesson count accurate?
- Any ADR number duplicates?
-
Doc freshness reverse-lint — catch stale normative guidance in project docs/ after this session's lessons.
The PostToolUse hook on Edit|Write already fires on memory-file edits, but this explicit pass catches:
- Lessons added outside a hooked Edit (e.g., via in-memory batch)
- Consolidated reports surfaced in the handoff doc itself
for memfile in $(git diff --name-only HEAD~N..HEAD | grep -E '(lessons|axioms|feedback_.*)\.md$'); do
python3 ~/.claude/skills/doc-freshness-reverse-lint/scripts/reverse_lint.py \
"$memfile" --project-root "$(pwd)" --human
done
Wire behavior:
- Zero candidates → exit silent, mention "reverse-lint: clean" in the summary table
- ≥1 candidate → add a "Stale docs to review" section to
session_N_handoff.md with
file:line references and the triggering rule. Never auto-edit the flagged docs; the
human decides what to update.
- If
reverse_lint.py is unavailable, log "doc-freshness-reverse-lint: not installed" and continue.
24b. Skill freshness audit (per axiom #21) — if any SKILL.md was edited this session, run the freshness check:
```bash
# Only if a SKILL.md was touched this session
if git diff --name-only HEAD~N..HEAD | grep -q 'skills/.*/SKILL\.md$'; then
# Resolve the bundled script — plugin install first, then git-clone install:
SFA="${CLAUDE_PLUGIN_ROOT:+${CLAUDE_PLUGIN_ROOT}/scripts/skill_freshness_audit.py}"
[ -f "$SFA" ] || SFA="$HOME/.claude/skills/session-handoff/scripts/skill_freshness_audit.py"
python3 "$SFA" --human
fi
```
Flags any skill whose `last_verified` has aged past `staleness_window_days` (default 90), or that opts into the freshness contract without declaring one. **Never auto-bump** `last_verified` — surface candidates for human verification and add them to the "Stale docs to review" section of the handoff doc.
Skip silently if no SKILL.md was touched. If the audit script is unavailable at both locations, log "skill_freshness_audit: not installed" and continue.
24c. Persist session usage metrics — archive this session's cctime output as a structured
record under ~/.claude/usage-tracking/ for cross-session analytics (Karpathy-style usage tracking).
**Canonical generator is the cctime FORK, invoked BY ABSOLUTE PATH** (not the bare
`cctime` name — see the name-collision warning below):
```bash
SID="${CLAUDE_CODE_SESSION_ID:?session id required}"
OUT="$HOME/.claude/usage-tracking/$(date -u +%Y-%m-%d)_${SID:0:8}_<project-short>"
CCTIME_FORK="$HOME/.claude/tools/cctime-fork/dist/index.js"
if [ -f "$CCTIME_FORK" ]; then
node "$CCTIME_FORK" --session "$SID" --json > "$OUT.json" # canonical record
node "$CCTIME_FORK" --session "$SID" > "$OUT.md" # human-readable companion
else
# Fork not present → in-skill recompute (same message.id dedup + recursive
# subagent accounting as the fork; tokens only, no cost math).
# Resolve the bundled script — plugin install first, then git-clone install:
SM="${CLAUDE_PLUGIN_ROOT:+${CLAUDE_PLUGIN_ROOT}/scripts/session_metrics.py}"
[ -f "$SM" ] || SM="$HOME/.claude/skills/session-handoff/scripts/session_metrics.py"
python3 "$SM" --session "$SID" --project=<slug-with-leading-dash> --json > "$OUT.json"
python3 "$SM" --session "$SID" --project=<slug-with-leading-dash> --print-summary > "$OUT.md"
fi
```
**CRITICAL — invoke the fork by PATH, never the bare `cctime`.** The fork's
`package.json` keeps the upstream package name `@dioptx/cctime`, so the global
`cctime` symlink is *ambiguous and unstable*: a `npm link` from the fork points it at
the fork, but any later `npm install -g @dioptx/cctime` (or a fresh machine) silently
clobbers it back to the **broken upstream**, and `cctime -V` reports `1.0.0` either way
so you can't tell them apart. Calling `node ~/.claude/tools/cctime-fork/dist/index.js`
sidesteps the symlink entirely and is deterministic. (This is the
`cctime-record-main-loop-inflated-by-stale-binary` failure mode — a stale-upstream
binary inflates main-loop cost ~1.8–2× via streaming-partial double-count, and the
wrong number is baked into the stored JSON.)
Why the fork at all: upstream (a) misattributes overnight-idle gaps to "Claude
thinking" (background-agent completion events during sleep classified as assistant
work), and (b) under-counts subagent tokens. Both are fixed in
[`wan-huiyan/cctime`](https://github.com/wan-huiyan/cctime) (PRs to upstream pending).
**One-time fork setup** (only if `~/.claude/tools/cctime-fork/dist/index.js` is absent):
```bash
git clone --depth 1 https://github.com/wan-huiyan/cctime ~/.claude/tools/cctime-fork
cd ~/.claude/tools/cctime-fork && npm install && npm run build
# No `npm link` needed — the skill calls dist/index.js by path.
```
Sanity check: on a session with an overnight gap, the fork reports "Xh away" as a
separate top-line beside "active"; upstream merges it into "Claude thinking" at ~98%.
**Fallback (`session_metrics.py`)** — if cctime isn't installed or the fork isn't available,
use the bundled Python recompute (resolve the path plugin-first, as above):
```bash
SM="${CLAUDE_PLUGIN_ROOT:+${CLAUDE_PLUGIN_ROOT}/scripts/session_metrics.py}"
[ -f "$SM" ] || SM="$HOME/.claude/skills/session-handoff/scripts/session_metrics.py"
python3 "$SM" --session "$SID" --project=<slug-with-leading-dash> --print-summary
```
This script dedupes tokens by `message.id` (whole-file, keeping the max-output chunk) and
sums main + recursively-discovered subagent JSONLs — the same accounting rules as the fork.
Tokens only: it does no time/idle-gap analysis and **no cost math** (pricing constants go
stale; cost estimation is delegated to the cctime fork / token-torch). Slower than cctime,
but works without it.
**Subagent token/cost accounting (now correct in BOTH tools as of 2026-05-29).** Subagent spend
(foreground Agent dispatches AND Workflow fan-outs) used to be undercounted; both the cctime fork
and `session_metrics.py` were fixed this session and now agree (~$59 / 17 transcripts on the
reference session). Three traps the fixes address — relevant if you write your own recompute:
- **Nesting:** Workflow subagents live one level deeper than foreground ones —
`<session>/subagents/agent-*.jsonl` (foreground) vs
`<session>/subagents/workflows/wf_<runid>/agent-*.jsonl` (workflow). RECURSE; a flat
`glob("agent-*.jsonl")` misses the fleet (tens of % of true cost).
- **Per-message dedup must keep the HIGHEST-output chunk, not the first.** Streaming chunks share
a message.id/requestId; the first carries full input/cache but output_tokens≈0, so first-wins
under-counts OUTPUT ~8x.
- **Dedup each request ONCE across the whole file, not just consecutive chunks.** An
order-preserving dedup that flushes its group on every non-assistant row (the cctime parser's
`deduplicateAssistant`) re-counts a request key that recurs across tool-result boundaries,
inflating subagent input/cache ~50%. cctime's NEW subagent path uses a whole-file dedup
(`src/subagents.ts`) instead of routing subagent files through the parser. (The main-loop path
still uses the order-preserving parser — its over-count is a minor ~1-3% and is left alone
because phase/time analysis needs message order.)
See skill `claude-code-workflow-subagent-tokens-nested-undercount`.
**Heads-up — the bundled `session_metrics.py` prints to stdout only** and never writes files
on its own, so it cannot clobber the cctime outputs. To persist the record, redirect its
`--json` output to `$OUT.json` and its `--print-summary` output to `$OUT.md` as shown in the
step-24c snippet above. Its record is tokens-only (no `$` figures) — if you need cost in the
archived record, install the cctime fork.
The step writes a pair of files: `<date>_<8-char-session-id>_<project>.{json,md}`. The JSON is the
canonical record (schema versioned, accumulatable via `jq -s` across sessions); the Markdown
is a human-readable companion. See `~/.claude/usage-tracking/README.md` for schema, methodology,
and cross-session query patterns.
**Wire behavior:**
- If `~/.claude/usage-tracking/README.md` doesn't exist, the script doesn't create it — surface
to the user that they should set up the tracking folder once.
- If neither cctime nor `session_metrics.py` is available, log "session metrics: not installed"
and continue.
- If both main and subagent JSONLs are missing for the session, skip silently — likely a
session that didn't go through the regular Claude Code transcript flow.
- On the FIRST run after a Claude Code update, sanity-check cctime against the Anthropic
billing dashboard — pricing constants may have moved.
25. Final confirmation to user: list all artifacts produced, grouped by bucket
Phase 5: Consolidate (when 3+ handoffs exist)
This phase runs automatically when 3+ handoff docs are detected, or when the user
explicitly asks to consolidate. It merges overlapping information from parallel
sessions into a single authoritative plan.
-
Gather all sources — read in parallel:
- All handoff docs (
docs/handoffs/session*_handoff.md)
- PR status (
gh pr list --state all --limit 20 --json number,title,state,mergedAt)
- Memory files (MEMORY.md, sessions_archive, lessons)
- Any status/findings docs
-
Track decision supersession — build a timeline across sessions:
-
Map experiments and identify gaps — cross-check every "What Needs To Happen Next"
section from every handoff against what actually happened:
- Promised work that was never started
- Planned validations that were skipped
- Integration items that were deferred and forgotten
- Branch cleanup that accumulated across sessions
-
Write consolidated plan -> docs/plans/future_sessions_plan.md
Structure:
# Consolidated Plan for Future Sessions
## Current State Summary
### Merged to main (PR table with outcomes)
### Open work (PRs, branches)
### Key findings (experiments, discoveries)
## Priority Actions (P1-PN)
(Each with: what, why, dependencies, experiment plan if applicable)
## Branch Cleanup (table with action per branch)
## Decision Queue
### Open decisions (numbered, with blockers)
### Resolved decisions (strikethrough, with rationale)
-
Verify all claims are current — for each claim in the consolidated plan:
- Is the PR status current? (
gh pr view N --json state)
- Are branch references still valid? (
git branch -a)
- Have any deferred items been completed without updating the plan?
Phase 6: User-facing live-dashboard recap (chat output)
After Phases 0-5 produce the persisted artefacts (handoff doc + buckets + PR),
deliver a separate, user-facing recap in chat that translates the
shipped work into what the user will actually see the next time they open
the product. This is conversational output, not a file — it's what the user
reads before they close the session.
Why this matters: handoff docs are written for Claude in a future session
(dense, technical, complete). The user reading the chat needs a different
register — what changed, where they'll notice it, anything they should
verify themselves. Without this step, the user has to read the handoff doc
or click through 4-8 PRs to know what changed in their product.
-
Structure the recap as "shipped change → user-visible effect", one
section per merged PR or material change. Skip purely internal work
(tracker entries, repo hygiene, doc-only PRs) — those don't surface to
the user.
For each change, write 2-4 lines covering:
- Where to look (which page / route / drawer / report)
- Before vs. after, framed in what the user actually perceived (not
class names, not SQL — what they SAW)
- Optional: any caveat (e.g. "no visible change but cleaner codebase")
-
Group by venue, not by PR, when multiple PRs land on the same page.
If three PRs all changed /drivers, write one /drivers recap section
summarising the net visible delta — not three sequential sections that
force the user to mentally compose.
-
Mention the baker / pipeline run if you ran one. "Baked payload
refreshed on <job-name> — the new labels are live now, not waiting
for the nightly bake." This closes the loop on "did the change
actually reach my eyes" — without it, the user wonders whether they're
looking at fresh data.
-
Flag pre-existing failures as pre-existing. If a test failed both
before and after your work, say so explicitly in the recap. Otherwise
the user assumes you introduced it.
-
Skip the recap when the work is purely internal. If the session was
e.g. memory-hygiene + skill edits + tracker reconciliation with zero
product-visible changes, just say so in one line: "Session was
internal-only — no live-dashboard impact." Don't manufacture user-
facing prose for backend hygiene.
Template (use as scaffolding; collapse sections that don't apply):
## 🌐 What you'll see in the live dashboard
### <Venue 1, e.g. `/drivers` chevron — `Days Since FAFSA` row>
- **Before:** <what the user used to see>
- **After:** <what the user sees now>
- <optional caveat>
### <Venue 2 — group multiple related PRs on the same page together>
...
### Baker / pipeline run *(only if you ran one)*
<job name> succeeded — <what's now live without waiting>.
### Notes
- <pre-existing failures you verified weren't introduced>
- <anything the user should sanity-check themselves>
Anti-patterns for the recap:
- ❌ Listing PR numbers + commit SHAs as the structure (that's the handoff doc's job)
- ❌ Quoting class names, function names, or SQL — user doesn't care
- ❌ "We shipped 4 PRs" with no per-PR what-they'll-see — useless
- ❌ Skipping the baker / pipeline run note when one was triggered
- ❌ Manufacturing user-visible prose for internal-only sessions
Good signal: the recap reads like release notes for someone who didn't
follow the session — not like a status update for someone who did.
Output format
Present a bucket-grouped summary table at the end. Leave rows blank ("—") for buckets
the session didn't touch — don't fabricate entries.
| Bucket / artifact | Status |
|---|
docs/decisions/ (ADRs) | N new (NNNN-NNNN), M updated — or "—" |
docs/runbooks/ | N new/updated — or "—" |
docs/analysis/ | N new/updated — or "—" |
docs/references/ | N new/updated — or "—" |
docs/reviews/ | N new/updated — or "—" |
docs/handoffs/ | session_N_handoff.md + session_N+1_prompt.md (+ parallel prompts if any) |
docs/deliverables/ | N new artifacts — or "—" |
docs/plans/future_sessions_plan.md | Updated / consolidated (if Phase 5) |
memory/lessons.md | N new (total: M) |
memory/sessions_archive.md | Updated — bucket footprint noted |
MEMORY.md index | Updated |
| Session usage record (step 24c) | REQUIRED — ~/.claude/usage-tracking/<date>_<sid8>_<project>.{json,md} written (cost $X, N subagents) or explicit "skipped: " |
| Doc-freshness reverse-lint | Clean / N candidates surfaced in handoff doc |
| PR | #N — merged / open for review |
| Git status | All committed and pushed |
The "Session usage record" row is NOT blankable — it is the forcing function for step 24c,
which is otherwise easy to drop (it's a nested step late in Phase 4, after the PR merge, and the
Phase 6 recap reads as the "finale" — the classic multi-agent-skill-silent-phase-compression
tail-drop). You must either cite the written record path or state "skipped: ". If you
reach this table and the cell is empty, go back and run step 24c before presenting.
Anti-patterns
- Don't concatenate handoff docs — the value of consolidation is resolving conflicts and surfacing gaps, not appending
- Don't assume handoff docs are current — check git/PR status for every claim
- Don't keep resolved decisions as open — clutters the decision queue
- Don't hardcode test counts or line counts — they go stale immediately; use "as of PR #N" instead
- Don't skip the lessons scan — debugging patterns are the most valuable long-term knowledge
- Don't write "see above" in next-session prompts — they must be paste-ready with full context
Tips
- Start with git log to jog memory about what happened
- Check reference files and architectural decisions for staleness
- If the session had a critical bug, add it to project CLAUDE.md (not just lessons)
- The next session prompt should be paste-ready — include all context needed to start immediately
- After consolidation, archive older handoff docs to reduce clutter
- The consolidated plan is a living document — update it when new sessions complete
Example output
A completed handoff produces this structure:
docs/handoffs/
session_12_handoff.md # What happened, what remains
session_13_prompt.md # Primary next-session prompt
session_13b_auth_cleanup_prompt.md # Parallel stream (zero file overlap)
docs/plans/
future_sessions_plan.md # (if Phase 5 ran) single source of truth
Parallel prompt example (session_13b_auth_cleanup_prompt.md):
# Session 13b — Auth Token Cleanup (Parallel)
**Branch:** `feat/s13b-auth-cleanup` (from main after PR #15 merge)
## Parallel session
S13 (rate limiting) runs on `feat/s13-rate-limiting`.
**No file overlap** — this session touches `auth/tokens/` only;
S13 touches `middleware/` + `config/`.
Handoff doc sections:
# Session 12 Handoff — Add user authentication flow
## Completed
- [x] OAuth2 integration (PR #15, 3 commits)
- [x] Token refresh middleware (42 tests pass)
## Remaining (prioritised)
1. **Rate limiting** — needs Redis config (ADR-0008)
2. **Session expiry UI** — stale mock data in fixtures
## Blockers & Open Issues
- **Redis not provisioned** — rate limiting blocked until infra team sets up Redis (asked in #ops-requests)
- **CI flake on auth tests** — intermittent timeout in `test_token_refresh`, not related to our changes
## Key Decisions
| Decision | Resolution | Rationale |
|---|---|---|
| Token storage | httpOnly cookies | XSS protection vs localStorage |
## Branch Status
- `feature/auth-flow` — 4 commits ahead of main, ready for PR
Decision supersession example (from consolidated plan):
## Decision Queue
1. **Should we add WebSocket support?** — OPEN, blocked on load testing results
2. ~~**Defer rate limiting?**~~ RESOLVED (Session 8). Implemented in PR #22 after abuse incident.
3. ~~**localStorage for tokens?**~~ SUPERSEDED (Session 5 reversed Session 3). httpOnly cookies chosen for XSS protection.
Composability
Input / Output Contract
Input: Invoked at end of a work session. Reads git log, memory files, and project structure.
For consolidation: reads all existing handoff docs and validates against git/GitHub state.
Output: Handoff doc, updated memory/lessons, next session prompt, sessions archive entry,
ADRs, and optionally a consolidated plan. All files are committed and pushed.
Dependencies
- Requires
git for commit history and status
- Requires
gh CLI for PR status checks (gracefully degrades without it)
- Works with any project structure that uses
docs/ and memory/ directories (creates them if missing)
- Optional (recommended):
doc-freshness-reverse-lint skill at
~/.claude/skills/doc-freshness-reverse-lint/scripts/reverse_lint.py — if absent, Phase 4 step 24
logs "not installed" and continues. Install from
https://github.com/wan-huiyan/claude-ecosystem-hygiene/tree/main/plugins/doc-freshness-reverse-lint.
- Taxonomy source of truth:
memory-hygiene v3.1+ defines the 7-bucket docs/ taxonomy.
Run memory-hygiene with --migrate if the target project's docs/ has drifted from the taxonomy
(loose files at docs/*, non-canonical subdirs, handoff/ + handoffs/ duplicates, etc.).
Scope Boundaries
- Use this skill when wrapping up a work session or consolidating after parallel sessions
- Do NOT use for mid-session progress updates (just use git commits)
- Hand off to
memory-hygiene for deep memory cleanup beyond what Phase 4 covers, and for
full docs/ taxonomy audits / migration branches
- Hand off to
doc-freshness-reverse-lint (invoked automatically in Phase 4 step 24) for
catching stale normative guidance in project docs after memory updates
